A transfer learning framework for cross-domain fault diagnosis of rotating machinery with imbalanced data

This article addresses the issues of sparse target-domain samples, class imbalance, and cross-domain distribution shifts in cross-condition fault diagnosis. It proposes a transfer learning framework that combines target-domain fault sample augmentation with class-level adversarial alignment. First, a variational autoencoder is used to train the model with abundant normal samples and a small number of fault samples from the target domain to generate pseudo-fault samples, thereby alleviating class imbalance in the target domain. Then, a discriminator is introduced for each class to achieve precise cross-domain distribution matching of samples within the same class, thereby reducing the risk of misalignment. We constructed 12 cross-operating-condition transfer tasks on the Case Western Reserve University and Wuhan University of Technology datasets. Experimental results achieved 91.91, 91.54, and 92.03% on the average accuracy, macro-F1, and balanced accuracy metrics, respectively, all outperforming the baseline methods. Through ablation experiments, experiments on unbalanced data strategies, and transfer experiments under different preprocessing conditions, the proposed method demonstrates excellent effectiveness, robustness, and cross-condition adaptability. This research provides an effective solution for cross-condition fault diagnosis of rotating machinery under sparse target data conditions.